Load scheduling optimization method and device, equipment and storage medium

Abstract

Embodiments of the present application provide a load scheduling optimization method, device, equipment and storage medium. The method first acquires power grid operation state data and user behavior data of at least one to-be-scheduled load node in a power distribution network, then performs feature extraction and attention weighting processing on the power grid operation state data and the user behavior data to obtain a load transfer potential evaluation value of the at least one to-be-scheduled load node, wherein the load transfer potential evaluation value is used to quantify the transferable capacity of the load node, and then a load scheduling strategy is determined according to the load transfer potential evaluation value of the at least one to-be-scheduled load node, wherein the load scheduling strategy includes a scheduling sequence and a scheduling action of each to-be-scheduled load node. By applying the technical solution of the present application, the transferable capacity of the load node can be accurately evaluated, and the load scheduling response speed is improved.

Description

Technical Field

[0001] This application relates to the field of smart grid and power system optimization scheduling technology, and in particular to a load scheduling optimization method, device, equipment and storage medium. Background Technology

[0002] In grid environments with high penetration of new energy sources, the proportion of renewable energy (such as photovoltaic and wind power) has increased significantly. However, their inherent intermittent and fluctuating characteristics exacerbate the difficulty of balancing grid supply and demand. For example, when photovoltaic output in an industrial park surges at midday, the grid needs to quickly transfer nearby loads to absorb the excess electricity and avoid equipment overload due to overvoltage. Conversely, when wind power output decreases at night, load transfer is necessary to fill the power supply gap. Therefore, load transfer scheduling is required to ensure grid supply and demand balance.

[0003] Existing transferable load scheduling schemes mainly rely on static optimization or simple dimensional evaluation, or adopt a two-level optimization architecture. The first level optimizes the load curve and battery scheduling strategy through particle swarm optimization to improve user satisfaction, while the second level optimizes equipment output based on a normal distribution model to reduce operating costs.

[0004] However, existing technologies are difficult to fully adapt to the rapid changes in new energy output and user demand, resulting in a weak load transfer and dispatch response capability. Summary of the Invention

[0005] This application provides a load scheduling optimization method, apparatus, equipment, and storage medium to improve the load transfer scheduling response capability.

[0006] In a first aspect, embodiments of this application provide a load scheduling optimization method, including:

[0007] Acquire grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network;

[0008] Feature extraction and attention weighting are performed on the power grid operation status data and the user behavior data to obtain the load transfer potential assessment value of the at least one load node to be scheduled. The load transfer potential assessment value is used to quantify the load node's transferability.

[0009] Based on the load transfer potential assessment value of the at least one load node to be scheduled, a load scheduling strategy is determined, the load scheduling strategy including: the scheduling order and scheduling action of each load node to be scheduled.

[0010] In one or more embodiments, the step of performing feature extraction and attention-weighted processing on the power grid operation status data and the user behavior data to obtain the load transfer potential assessment value of the at least one load node to be scheduled includes:

[0011] A preset spatiotemporal graph neural network model is used to extract spatial features from the power grid operation status data and the user behavior data to obtain spatial coupling feature data. The preset spatiotemporal graph neural network model includes: a feature extraction unit for extracting spatial correlation features and a weight allocation unit for dynamically adjusting the weights of spatial correlation features.

[0012] A multi-branch long short-term memory autoencoder is used to extract time-series features from the power grid operation status data and the user behavior data to obtain time-series feature data. The multi-branch long short-term memory autoencoder includes: multiple parallel long short-term memory encoders for extracting time-series features of different data respectively and a shared long short-term memory decoder for reconstructing time-series features.

[0013] The spatial coupling feature data and the temporal feature data are spliced ​​and fused to obtain spatiotemporal feature data;

[0014] Attention-weighted processing is applied to the spatiotemporal feature data to determine the load transfer potential assessment value.

[0015] In one or more embodiments, determining the load scheduling strategy based on the load transfer potential assessment value of the at least one load node to be scheduled includes:

[0016] Based on the power grid operation status data and preset scheduling actions, a reinforcement learning reward function is constructed, which is used to quantify the ratio of transfer benefits to transfer costs.

[0017] Based on the load transfer potential assessment value of the at least one load node to be scheduled, the reinforcement learning reward function is solved using the Monte Carlo tree search algorithm to obtain the load transfer value of the at least one load node to be scheduled.

[0018] The load dispatching strategy is determined based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers.

[0019] In one or more embodiments, determining the load dispatching strategy based on the power grid operating status data, all load transfer potential assessment values, and all load transfer values ​​includes:

[0020] Based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers, a mixed integer nonlinear programming method is used to determine the objective optimization function, which takes operating cost, carbon emissions, and load transfer revenue as optimization objectives.

[0021] The objective optimization function is solved using the ε-constraint method to obtain the load scheduling strategy.

[0022] In one or more embodiments, before determining the load scheduling strategy based on the load transfer potential assessment value of the at least one load node to be scheduled, the method further includes:

[0023] Determine the geographical coordinates of each load node to be scheduled and the similarity of electricity consumption patterns between any two load nodes to be scheduled;

[0024] The first load nodes to be dispatched that meet the preset conditions are clustered and aggregated into the first virtual power plant. The preset conditions are: the geographical coordinates are less than the preset distance threshold and the similarity of the electricity consumption change patterns is higher than the set similarity threshold.

[0025] The second load node that does not meet the preset conditions will be treated as a second virtual power plant.

[0026] Replace all load nodes to be scheduled based on the first virtual power plant and the second virtual power plant.

[0027] In one or more embodiments, the method further includes:

[0028] Based on the load scheduling strategy, a reward and penalty function that balances scheduling response speed and operational economy is constructed.

[0029] The load scheduling strategy is modified and adjusted using a near-end strategy optimization algorithm to obtain the adjusted load scheduling strategy.

[0030] In one or more embodiments, the method further includes:

[0031] The system uses blockchain smart contracts to automatically distribute economic compensation to at least one load node to be scheduled, and employs federated learning to protect the user behavior data.

[0032] Secondly, embodiments of this application provide a load scheduling optimization device, comprising:

[0033] The acquisition module acquires power grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network;

[0034] The processing module is used to perform feature extraction and attention weighting processing on the power grid operation status data and the user behavior data to obtain the load transfer potential assessment value of the at least one load node to be scheduled. The load transfer potential assessment value is used to quantify the transferability of the load node.

[0035] The determination module is used to determine a load scheduling strategy based on the load transfer potential assessment value of the at least one load node to be scheduled. The load scheduling strategy includes the scheduling order and scheduling action of each load node to be scheduled.

[0036] In one or more embodiments, the processing module is specifically used for:

[0037] A preset spatiotemporal graph neural network model is used to extract spatial features from the power grid operation status data and the user behavior data to obtain spatial coupling feature data. The preset spatiotemporal graph neural network model includes: a feature extraction unit for extracting spatial correlation features and a weight allocation unit for dynamically adjusting the weights of spatial correlation features.

[0038] A multi-branch long short-term memory autoencoder is used to extract time-series features from the power grid operation status data and the user behavior data to obtain time-series feature data. The multi-branch long short-term memory autoencoder includes: multiple parallel long short-term memory encoders for extracting time-series features of different data respectively and a shared long short-term memory decoder for reconstructing time-series features.

[0039] The spatial coupling feature data and the temporal feature data are spliced ​​and fused to obtain spatiotemporal feature data;

[0040] Attention-weighted processing is applied to the spatiotemporal feature data to determine the load transfer potential assessment value.

[0041] In one or more embodiments, the determining module is specifically used for:

[0042] Based on the power grid operation status data and preset scheduling actions, a reinforcement learning reward function is constructed, which is used to quantify the ratio of transfer benefits to transfer costs.

[0043] Based on the load transfer potential assessment value of the at least one load node to be scheduled, the reinforcement learning reward function is solved using the Monte Carlo tree search algorithm to obtain the load transfer value of the at least one load node to be scheduled.

[0044] The load dispatching strategy is determined based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers.

[0045] In one or more embodiments, the determining module determines the load dispatching strategy based on the power grid operating status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers, specifically for:

[0046] Based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers, a mixed integer nonlinear programming method is used to determine the objective optimization function, which takes operating cost, carbon emissions, and load transfer revenue as optimization objectives.

[0047] The objective optimization function is solved using the ε-constraint method to obtain the load scheduling strategy.

[0048] In one or more embodiments, before determining the load scheduling strategy based on the load transfer potential assessment value of the at least one load node to be scheduled, the determining module is further configured to:

[0049] Determine the geographical coordinates of each load node to be scheduled and the similarity of electricity consumption patterns between any two load nodes to be scheduled;

[0050] The first load nodes to be dispatched that meet the preset conditions are clustered and aggregated into the first virtual power plant. The preset conditions are: the geographical coordinates are less than the preset distance threshold and the similarity of the electricity consumption change patterns is higher than the set similarity threshold.

[0051] The second load node that does not meet the preset conditions will be treated as a second virtual power plant.

[0052] Replace all load nodes to be scheduled based on the first virtual power plant and the second virtual power plant.

[0053] In one or more embodiments, the determining module is further configured to:

[0054] Based on the load scheduling strategy, a reward and penalty function that balances scheduling response speed and operational economy is constructed.

[0055] The load scheduling strategy is modified and adjusted using a near-end strategy optimization algorithm to obtain the adjusted load scheduling strategy.

[0056] In one or more embodiments, the determining module is further configured to:

[0057] The system uses blockchain smart contracts to automatically distribute economic compensation to at least one load node to be scheduled, and employs federated learning to protect the user behavior data.

[0058] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0059] The memory stores computer-executed instructions;

[0060] The processor executes computer execution instructions stored in the memory, such that the processor, when executed, is used to implement the method described in the first aspect and any of the embodiments above.

[0061] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods described in the first aspect and any of the embodiments above.

[0062] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, is used to implement the load scheduling optimization method as described in the first aspect and various possible implementations of the first aspect.

[0063] This application provides a load dispatch optimization method, apparatus, device, and storage medium. The method first acquires grid operation status data and user behavior data of at least one load node to be dispatched in a distribution network. Then, it performs feature extraction and attention weighting processing on the grid operation status data and user behavior data to obtain a load transfer potential assessment value for at least one load node to be dispatched. This load transfer potential assessment value is used to quantify the load node's transferability. Finally, based on the load transfer potential assessment values ​​of at least one load node to be dispatched, a load dispatch strategy is determined. This load dispatch strategy includes the dispatch sequence and dispatch actions for each load node to be dispatched. In the above method, by acquiring the grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network, and performing feature extraction and attention weighting processing on the grid operation status data and user behavior data, a load transfer potential assessment value is obtained to quantify the load transfer capability of the load node. This can accurately quantify the load transfer characteristics of each load node to be dispatched. Then, based on the load transfer potential assessment value of at least one load node to be dispatched, a load dispatching strategy including the dispatching sequence and dispatching actions of each load node to be dispatched is determined. This can improve the response capability of load coordinated dispatching to fluctuations in new energy output and changes in user demand, thereby improving the overall efficiency and optimization level of distribution network load dispatching. Attached Figure Description

[0064] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0065] Figure 1 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 1 ;

[0066] Figure 2 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 2 ;

[0067] Figure 3 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 3 ;

[0068] Figure 4 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 4 ;

[0069] Figure 5 This is a schematic diagram of the load scheduling optimization device provided in the embodiments of this application;

[0070] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0071] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0072] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0073] Before introducing the embodiments of this application, the application background of the embodiments of this application will be explained first:

[0074] In grid environments with high penetration of new energy sources, the proportion of renewable energy (such as photovoltaic and wind power) has increased significantly. However, their inherent intermittent and fluctuating characteristics exacerbate the difficulty of balancing grid supply and demand. For example, when photovoltaic output in an industrial park surges at midday, the grid needs to quickly transfer nearby loads to absorb the excess electricity and avoid equipment overload due to overvoltage. Conversely, when wind power output decreases at night, load transfer is necessary to fill the power supply gap. Therefore, load transfer scheduling is required to ensure grid supply and demand balance.

[0075] Existing transferable load scheduling schemes mainly rely on static optimization or simple dimensional evaluation, or adopt a two-level optimization architecture. The first level optimizes the load curve and battery scheduling strategy through particle swarm optimization to improve user satisfaction, while the second level optimizes equipment output based on a normal distribution model to reduce operating costs.

[0076] However, existing technologies are difficult to fully adapt to the rapid changes in new energy output and user demand, resulting in a weak load transfer and dispatch response capability.

[0077] The load dispatch optimization method provided in this application aims to solve the aforementioned technical problems of the prior art. The technical concept of this application is as follows: Existing methods suffer from weak dynamic response capability in the coordinated dispatch of transferable loads under scenarios of fluctuating renewable energy output. To address the technical problems of existing methods, if the operational data of dispatchable load nodes in the distribution network can be jointly analyzed, and dispatchable load nodes with higher transfer potential are prioritized, the response capability of load transfer dispatch can be effectively improved. Therefore, the embodiments of this application consider acquiring grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network, and performing feature extraction and attention weighting processing on the grid operation status data and user behavior data to obtain a load transfer potential assessment value used to quantify the transferability of load nodes. Then, based on the load transfer potential assessment value of at least one load node to be dispatched, a load dispatch strategy including the dispatch order and dispatch actions of each load node to be dispatched is determined, thereby improving the response capability of the load dispatch strategy to fluctuations in renewable energy output and changes in user demand.

[0078] The execution subject of this application embodiment is an electronic device, which can be a terminal device, such as a laptop, desktop computer, or tablet computer, or a server. In practical applications, whether the electronic device is a terminal device or a server can be determined according to the actual situation, and no specific limitation is imposed on it.

[0079] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0080] Figure 1 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 1 .like Figure 1 As shown, the load scheduling optimization method includes the following steps:

[0081] S101. Obtain grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network.

[0082] In this step, the distribution network is a distribution network with a high proportion of new energy access, which connects distributed photovoltaic, wind power and user-side adjustable loads. The load nodes to be dispatched are the transferable load units in the distribution network. In order to optimize load dispatch, the grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network can be obtained first to provide data basis for subsequent load dispatch optimization.

[0083] For example, power grid operation status data is used to reflect the operation status of load nodes to be dispatched in the distribution network at the current time or within a preset time window, including: load curve, node voltage, active power, reactive power, branch power flow, and one or more of new energy output.

[0084] User behavior data is used to describe the electricity consumption behavior characteristics of users corresponding to the load nodes to be scheduled, including one or more of the following: historical electricity consumption change patterns, electricity consumption time distribution, load start-up and shutdown habits, and continuous running duration.

[0085] In one possible implementation, the method execution entity of this application embodiment can be deployed in a distribution automation master station, a park energy management platform, an edge computing gateway, or a scheduling server, and connected to the distribution network operation monitoring unit and the user-side energy consumption acquisition unit respectively through standard communication interfaces. The distribution network operation monitoring unit uploads real-time measurement data according to a set sampling period; the user-side energy consumption acquisition unit collects user behavior data through smart meters, load control terminals, charging pile controllers, building management system interfaces, or industrial control gateways.

[0086] After obtaining the grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network, missing and outlier values ​​can be filled and corrected. For example, for missing values, interpolation of nearby times, filling with the mean of the same type of node, or filling with the most recent valid value can be used to form a complete input sample. For outlier values, they are identified based on physical boundary conditions and rate of change thresholds. Samples that exceed the rated capacity limit, have abnormal negative power direction, or have short-term sudden changes exceeding a preset multiple are marked and corrected.

[0087] In addition, real-time weather forecast data and economic operation data can be obtained to consider the impact of weather and economy on the transfer potential of load nodes. Among them, weather forecast data directly determines the electricity demand of flexible loads such as distributed photovoltaic power output and air conditioning / heating, and is the core driving factor of load fluctuation and new energy output. For example, high irradiance on sunny days will generate a large amount of surplus electricity, requiring an increase in load transfer space to absorb the electricity, while high temperatures and severe cold will significantly increase temperature control loads and compress the upper limit of adjustable loads. Economic operation data can reflect the activity level of industrial and commercial production and shop operations in the distribution area, directly determining the scale of industrial transferable load and the pattern of electricity consumption. For example, when economic operation data is high, there are more industrial adjustable devices and greater load transfer potential.

[0088] S102. Perform feature extraction and attention weighting processing on power grid operation status data and user behavior data to obtain a load transfer potential assessment value for at least one load node to be scheduled.

[0089] Among them, the load transfer potential assessment value is used to quantify the load transfer capability of load nodes.

[0090] In this step, feature extraction is performed on power grid operation status data and user behavior data to extract key information that reflects the load transfer capability of load nodes. Attention weighting is then applied to this key information, assigning different importance to different information, thereby obtaining a load transfer potential assessment value used to quantify the load transfer capability of load nodes.

[0091] For example, the load transfer potential assessment value is a quantitative result output for each load node to be scheduled. It can be represented by a normalized score, a percentage score or a discrete level value between 0 and 1, which is used to characterize the ability of the load node to be scheduled to be transferred, reduced, postponed or recovered in the current scheduling cycle.

[0092] In one possible implementation, the power grid operation status data and user behavior data can be preprocessed, including normalization, discrete feature encoding, and time window slicing. Normalization can employ minimum-maximum mapping or Z-score standardization to ensure that data with different dimensions can be jointly modeled. Discrete feature encoding is used to transform behavioral information such as load type, shift type, and workday attributes into a computable representation. Time window slicing forms a historical observation sequence based on multiple consecutive sampling points.

[0093] Next, features that characterize the transferability of load nodes to be dispatched are extracted based on the preprocessed power grid operation status data and user behavior data. The extracted features may include relevant features reflecting the node operation status, historical change patterns and user energy consumption behavior. The extracted features are then comprehensively processed to form a comprehensive characterization of each load node to be dispatched.

[0094] The comprehensive characterization is then subjected to attention-weighted processing to highlight factors that have a more significant impact on load transfer potential, such as voltage over-limit risk, branch load rate increases, instantaneous output fluctuations of renewable energy sources, user fixed production period constraints, and continuous load operating duration. Overall weighting, partial weighting, or hierarchical weighting methods can be used to differentiate between different characteristics and their degree of influence, thereby improving the sensitivity of the assessment results to changes in current operating conditions and user behavior.

[0095] Finally, the load transfer potential assessment value is obtained based on the weighted feature calculation. During the execution phase, real-time power grid operation status data and user behavior data are input for the current scheduling cycle, and a list of load transfer potential assessment values ​​corresponding to each load node to be scheduled is output.

[0096] S103. Determine the load scheduling strategy based on the load transfer potential assessment value of at least one load node to be scheduled;

[0097] The load scheduling strategy includes the scheduling order and scheduling actions of each load node to be scheduled.

[0098] In this step, based on the load transfer potential assessment value of at least one load node to be scheduled, the transfer capability of each load node to be scheduled is determined. Each load node to be scheduled is sorted according to the size of its transfer capability, and the scheduling action of each load node is determined, resulting in a load scheduling strategy that includes the scheduling order and scheduling action of each load node to be scheduled.

[0099] For example, the scheduling sequence is used to determine the order in which each load node to be scheduled participates in load scheduling, and the scheduling action is used to specify the control operation to be performed on each load node to be scheduled. The action type includes one or more of the following: load transfer, load reduction, load restoration, power consumption adjustment, and power consumption time change.

[0100] In one possible implementation, a set of candidate scheduling nodes is selected based on the load transfer potential assessment value of each load node to be scheduled, combined with execution constraints. The execution constraints include at least the upper and lower limits of the node's rated power, the shortest start-stop interval, the minimum continuous operation duration, the maximum allowable delay duration, the user's uninterrupted time period, voltage constraints, feeder capacity constraints, and transformer capacity constraints.

[0101] For load nodes that do not meet the basic execution conditions, such as those in uninterrupted production periods, those that have reached the minimum continuous operating time limit, or those that are not allowed to be further reduced or restored in the current distribution area, they can be removed from the candidate scheduling node set or have their priority reduced. The candidate scheduling nodes are then sorted from high to low according to their load transfer potential assessment value to form an initial scheduling sequence. If there are multiple load transfer potential assessment values ​​with the same value, they can be further sorted according to the power of the load nodes to be scheduled, the tension of their respective feeders, response delay, and user behavior stability.

[0102] Scheduling actions are generated through a combination of rules and optimization. The rule part determines the set of candidate actions based on the candidate scheduling node type and constraints, such as allowing charging station nodes to perform charging delays, charging power reductions, and time-sharing recovery. The optimization part searches the candidate action set for action combinations that meet the current distribution network operation requirements.

[0103] If the output of renewable energy sources suddenly increases during the current period and local feeders have capacity to absorb the load, load transfer to the current period or load restoration actions can be issued to high-potential nodes. If the output of renewable energy sources suddenly drops, the power flow of branches approaches its upper limit, or the node voltage is low, load reduction, power reduction, or time shift actions will be prioritized for high-potential nodes. Each action is configured with an action start time, duration, adjusted power amplitude, and recovery conditions, forming an executable scheduling command used to control the corresponding node to be scheduled to perform the scheduling action.

[0104] In one possible implementation, prior to step S103 above, the load scheduling optimization method further includes the following steps:

[0105] Step 1: Determine the geographical coordinates of each load node to be scheduled and the similarity of electricity consumption patterns between any two load nodes to be scheduled.

[0106] For example, obtain the latitude and longitude coordinates or distribution network topology coordinates corresponding to each load node to be scheduled, and calculate the similarity between the spatial distance and the electricity consumption change pattern for any two load nodes to be scheduled.

[0107] In one possible implementation, the similarity of electricity consumption patterns can be determined by first using the load time-series data in the power grid operation status data of the two load nodes to be scheduled, and then calculating the degree of matching of electricity consumption fluctuation patterns using a cosine similarity algorithm.

[0108] Step 2: Cluster the first load nodes to be dispatched that meet the preset conditions into the first virtual power plant;

[0109] The preset conditions are: the geographical coordinates are less than a preset distance threshold and the similarity of the electricity consumption change pattern is higher than a set similarity threshold.

[0110] For example, when the spatial distance between two load nodes to be scheduled is less than a preset distance threshold and the similarity of their electricity consumption patterns is higher than a set similarity threshold, they are included in the same cluster set, and the cluster set is mapped to the first virtual power plant.

[0111] In one possible implementation, the first virtual power plant can consist of the aggregated load curve, aggregated adjustable capacity, and unified dispatch identifier of the first load node to be dispatched that meets preset conditions. It also aggregates the flexible load, distributed photovoltaic and energy storage resources of all the first load nodes to be dispatched and uniformly calculates the overall adjustable load limit of the first virtual power plant.

[0112] Step 3: Treat the second load node that does not meet the preset conditions as a second virtual power plant.

[0113] For example, when a load node to be scheduled cannot simultaneously meet the preset conditions with other load nodes to be scheduled, it is treated as an isolated load node that cannot be clustered and aggregated, i.e., a second load node to be scheduled that does not meet the preset conditions, and is treated as a second virtual power plant. The second virtual power plant retains the original adjustable attributes and scheduling identifier of the load node to be scheduled.

[0114] Step 4: Replace all load nodes to be dispatched based on the first and second virtual power plants.

[0115] For example, aggregated first virtual power plants and independent second virtual power plants can replace the original discrete load nodes to be dispatched in subsequent optimization calculations.

[0116] In this embodiment, the processing method combines geographical coordinates and similarity of electricity consumption change patterns for joint screening, and uniformly replaces discrete load nodes to be dispatched with two types of virtual power plants: aggregated and independent. Subsequent load dispatch optimization uses virtual power plants as the smallest dispatch unit, which greatly reduces the number of variables and improves the solution speed.

[0117] The load dispatch optimization method provided in this application first obtains the power grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network. Then, it performs feature extraction and attention weighting processing on the power grid operation status data and user behavior data to obtain the load transfer potential assessment value of at least one load node to be dispatched. The load transfer potential assessment value is used to quantify the transferability of the load node. Then, based on the load transfer potential assessment value of at least one load node to be dispatched, a load dispatch strategy is determined. The load dispatch strategy includes the dispatch order and dispatch action of each load node to be dispatched. In this embodiment, by acquiring grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network, and performing feature extraction and attention weighting processing on the grid operation status data and user behavior data, a load transfer potential assessment value for quantifying the load node's transferability is obtained. This can accurately quantify the load transfer characteristics of each load node to be dispatched. Then, based on the load transfer potential assessment value of at least one load node to be dispatched, a load dispatching strategy including the dispatching order and dispatching actions of each load node to be dispatched is determined. This can improve the response capability of load coordinated dispatching to fluctuations in new energy output and changes in user demand, thereby improving the overall efficiency and optimization level of distribution network load dispatching.

[0118] Based on the above embodiments, Figure 2 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 2 .like Figure 2 As shown, one possible implementation of step S102 above includes the following steps:

[0119] S201. A preset spatiotemporal neural network model is used to extract spatial features from power grid operation status data and user behavior data to obtain spatial coupling feature data.

[0120] The preset spatiotemporal graph neural network model includes: a feature extraction unit for extracting spatial correlation features and a weight allocation unit for dynamically adjusting the weights of spatial correlation features.

[0121] In this step, the power grid operation status data and user behavior data are extracted by the feature extraction unit in the preset spatiotemporal diagram neural network model to extract spatial correlation features. The weights of the spatial correlation features are dynamically adjusted by the weight allocation unit in the preset spatiotemporal diagram neural network model to obtain spatial coupling feature data that reflects the adjustable load linkage law of the distribution network spatial dimension.

[0122] In one possible implementation, the pre-defined spatiotemporal graph neural network model can use graph convolution as the feature extraction unit and attention layer as the weight allocation unit.

[0123] First, a graph structure containing multiple load nodes to be dispatched and the connection relationships between nodes is constructed based on the power grid operation status data and the distribution network topology. Then, a topological adjacency matrix is ​​constructed, and spatial correlation modeling is performed on the power grid operation status data, user behavior data, real-time weather forecast data, and economic operation data based on the adjacency matrix to obtain spatial correlation feature data.

[0124] Next, the attention mechanism of the attention layer is used to quantify the spatial association weights between different nodes to be scheduled in the spatial association feature data, and output the spatial coupling feature vector.

[0125] Furthermore, the model parameters of the preset spatiotemporal graph neural network model can be obtained by training with labeled sample data of historical scheduling nodes. In practical applications, other neural network models can also be selected, and this application does not limit this.

[0126] S202. A multi-branch long short-term memory autoencoder is used to extract time-series features from power grid operation status data and user behavior data to obtain time-series feature data.

[0127] The multi-branch long short-term memory autoencoder includes: multiple parallel long short-term memory encoders for extracting different data temporal features respectively, and a shared long short-term memory decoder for reconstructing temporal features.

[0128] In this step, multiple parallel long short-term memory encoders in a multi-branch long short-term memory autoencoder are used to independently parse the categorized time-series features of power grid operation status data and user behavior data. Furthermore, a shared long short-term memory decoder in the multi-branch long short-term memory autoencoder is used to denoise and reconstruct the categorized time-series features to obtain time-series feature data.

[0129] In one possible implementation, multiple independent parallel long short-term memory encoder branches are built, corresponding to four types of data: power grid operation status data, user behavior data, real-time weather forecast data, and economic operation data, respectively. All branches output classified time-series features, which are then concatenated and input into a single-layer shared long short-term memory decoder. Data noise is removed by reconstructing the loss function, and multi-channel time-series feature data is output.

[0130] S203. The spatial coupling feature data and the temporal feature data are spliced ​​and fused to obtain spatiotemporal feature data.

[0131] In this step, spatially coupled feature data and temporal feature data are spliced ​​and fused along the feature dimension to fully preserve all information of both spatial correlation and temporal fluctuation, thus obtaining spatiotemporal feature data, which provides a complete standardized input for attention weighting.

[0132] In one possible implementation, a splicing operation is performed on the feature channel dimension without dimensionality reduction compression. The independent channels of spatially coupled feature data and temporal feature data of each branch are preserved, and a long vector of uniform dimension is generated as the fused spatiotemporal feature data.

[0133] S204. Perform attention-weighted processing on the spatiotemporal characteristic data to determine the load transfer potential assessment value.

[0134] In this step, a grouped multi-head attention mechanism is adopted to automatically assign the influence weights of different factors in the spatiotemporal feature data, quantify the degree of influence of each factor on the adjustable load, and perform regression on the attention-weighted spatiotemporal feature data to obtain quantified values ​​as load transfer potential assessment values.

[0135] In one possible implementation, a grouped multi-head attention module is first used to weight the spatiotemporal feature data sources. For example, global attention processing is performed on the spatiotemporal feature data to obtain the overall influence weights of user behavior factors and meteorological factors. Local attention processing is then performed on the spatiotemporal feature data that has undergone global attention processing to obtain the dynamic influence weights of new energy output fluctuation factors. Finally, the influence weights of different factors on load transfer are generated based on the overall influence weights and dynamic influence weights, and the spatiotemporal feature data sources are weighted according to the influence weights of different factors on load transfer. Then, the weighted spatiotemporal feature data is connected to a fully connected regression network model, and the fully connected regression network model is trained with the historical transferable load ground values. The regression output is the load transfer potential assessment value of a single load node to be scheduled.

[0136] The load dispatch optimization method provided in this application first uses a preset spatiotemporal graph neural network model to extract spatial features from power grid operation status data and user behavior data to obtain spatially coupled feature data. The preset spatiotemporal graph neural network model includes a feature extraction unit for extracting spatial correlation features and a weight allocation unit for dynamically adjusting the weights of spatial correlation features. Next, a multi-branch long short-term memory autoencoder is used to extract temporal features from power grid operation status data and user behavior data to obtain temporal feature data. The multi-branch long short-term memory autoencoder includes multiple parallel long short-term memory encoders for extracting temporal features of different data respectively and a shared long short-term memory decoder for reconstructing temporal features. Then, the spatially coupled feature data and temporal feature data are spliced ​​and fused to obtain spatiotemporal feature data. Finally, attention weighting processing is applied to the spatiotemporal feature data to determine the load transfer potential assessment value. In this embodiment, after spatial modeling and temporal modeling are executed in parallel, the attention-weighted result is used as the basis for calculating the load transfer potential assessment value and is associated with the load nodes to be scheduled. This allows each load node to be scheduled to be quantitatively scored for its transfer potential, effectively reflecting the electrical coupling relationship and load change pattern between the load nodes to be scheduled, and providing real-time and accurate assessment of the load transfer potential.

[0137] Based on the above embodiments, Figure 3 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 3 .like Figure 3 As shown, one possible implementation of step S103 above includes the following steps:

[0138] S301. Construct a reinforcement learning reward function based on power grid operation status data and preset scheduling actions;

[0139] Among them, the reinforcement learning reward function is used to quantify the ratio of transfer benefits to transfer costs.

[0140] In this step, a reinforcement learning reward function is defined based on power grid operation status data to quantify the ratio of transfer benefits to transfer costs. This function uniformly measures the overall economic and environmental cost-effectiveness of preset dispatch actions, providing a unified evaluation standard for subsequent dispatch optimization.

[0141] For example, preset scheduling actions may include: load transfer, load reduction, load restoration, power consumption adjustment, and power consumption period change.

[0142] In one possible implementation, a reinforcement learning framework is constructed, defining a state-action-reward space:

[0143] The state space includes: power grid operation status data, user behavior data, real-time weather forecast data, and economic operation data;

[0144] Action space: Pre-set scheduling actions that meet power grid security constraints;

[0145] Reward function: Define the "transfer benefit-cost ratio" metric. The calculation formula is expressed as follows:

[0146]

[0147] in, The benefits of load transfer include: changes in operating costs before and after the load transfer. Carbon emission reduction benefits before and after load shift and the benefits of reduced grid losses before and after load transfer. , This represents the switching costs, including: user compensation costs. and equipment depreciation costs .

[0148] The numerator integrates the reduced operating costs, carbon emission reduction benefits, and reduced grid losses resulting from load transfer; the denominator counts two types of control costs: user compensation and equipment losses. The higher the ratio of the reward function output, the better the overall benefit of the preset adjustment action. Furthermore, grid electricity price, carbon price, and equipment depreciation parameters can be substituted into the reward function to complete the initialization.

[0149] S302. Based on the load transfer potential assessment value of at least one load node to be scheduled, the Monte Carlo tree search algorithm is used to solve the reinforcement learning reward function to obtain the load transfer value of at least one load node to be scheduled.

[0150] In this step, the load transfer potential assessment value of at least one load node to be scheduled is used as the hard constraint of the adjustment capacity of the load node to be scheduled. The Monte Carlo tree search algorithm is used to solve the reinforcement learning reward function. During the search process, simulation, backpropagation and updating are carried out for different scheduling orders and different scheduling actions. The average reward of each preset scheduling action is iteratively updated, and finally the comprehensive quantitative load transfer value of each load node to be scheduled is obtained by convergence.

[0151] In one possible implementation, a Monte Carlo search tree is built with a single load node to be scheduled as the root node. The UCB formula of the Monte Carlo tree search algorithm is used to filter branches. The node expansion, multi-scenario random simulation, and reward backtracking update loop are executed in sequence. The load transfer potential evaluation value is set as the upper limit of the action. After multiple iterations, the average reward of each load node to be scheduled is taken as the load transfer value.

[0152] The UCB formula is expressed as follows:

[0153]

[0154] in, Indicates the first The historical average reward of a pre-defined scheduled action. This represents the total number of simulations. Indicates the first The number of simulations for a preset scheduled action. This represents the exploration coefficient, balancing the use of known high-yield actions with the exploration of potential high-yield actions.

[0155] During the node expansion phase, when a leaf node is reached, new child nodes are expanded; during the multi-scenario random simulation phase, action sequences are randomly simulated starting from the new child nodes to predict possible reward values; during the reward backtracking phase, the simulated reward values ​​are backpropagated to the parent nodes to update the average reward and number of visits for each node.

[0156] S303. Determine the load dispatch strategy based on the power grid operation status data, the assessment value of all load transfer potential, and the value of all load transfers.

[0157] In this step, load transfer value is used as the core sorting criterion. The load transfer potential assessment value limits the maximum regulation capacity of each load node to be regulated. The scheduling order and the regulation amount of each load node to be regulated are divided in combination with the real-time grid load and voltage safety constraints in the grid operation status data, and a load scheduling strategy including scheduling order and scheduling actions is generated.

[0158] In one possible implementation, all load nodes to be scheduled are sorted from high to low according to their load transfer value, and the adjustment capacity is preferentially allocated to high-value nodes. The load transfer potential value is used as the adjustment limit for each load node to be scheduled. Under the safety constraints of distribution network voltage or line current carrying capacity, the priority sequence of global load scheduling and the corresponding scheduling actions are output to determine the load scheduling strategy.

[0159] In one possible implementation, step S303 described above may include the following steps:

[0160] Step 1: Based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers, perform mixed integer nonlinear programming to determine the objective optimization function;

[0161] The objective optimization function takes operating cost, carbon emissions, and load transfer benefits as optimization objectives.

[0162] For example, based on the power grid operation status data, the assessment values ​​of all load transfer potential, and the value of all load transfers, a three-objective optimization function of Mixed-Integer Nonlinear Programming (MINLP) is constructed, which simultaneously takes into account the three optimization objectives of operating cost, carbon emissions, and load transfer benefits.

[0163] In one possible implementation, total operating cost and total carbon emissions are set as minimization terms, and comprehensive load transfer benefits are set as maximization terms. Normalized weights are introduced to balance the emphasis of the three. At the same time, the load transfer potential assessment value is used as the upper limit of the adjustment amount of each node to be adjusted, and the load transfer value is used as the scheduling priority preference. Furthermore, energy storage, distribution network security, and carbon trading rules are introduced to form complete constraints, and linear, nonlinear, and discrete variables are unified and integrated to form an objective optimization function adapted to the virtual power plant.

[0164] For example, the objective function in the form of a weighted summation can be represented as follows:

[0165]

[0166] in, This represents the total operating cost. This represents total carbon emissions. Indicates the transfer of income. , ,as well as These represent the weight coefficients, set using the analytic hierarchy process or user preferences, and satisfy the following conditions: .

[0167] The constraints of the objective optimization function include: the total load transfer volume of each virtual power plant does not exceed the aggregated load transfer potential assessment value, and high transfer value nodes are given priority in the allocation of regulation capacity.

[0168] Step 2: Solve the objective optimization function using the ε-constraint method to obtain the load scheduling strategy.

[0169] For example, the ε-constraint method is used to handle the multi-objective conflict problem of the objective optimization function. Load transfer revenue is selected as the main optimization objective, and operating costs and carbon emissions are transformed into ε-constraint boundaries. Multiple Pareto optimal solutions are generated by iteratively adjusting the ε threshold, and a balanced scheme that takes into account economic, environmental and regulatory benefits is selected from them. After the solution is completed, the time-sharing energy storage charging and discharging power of each virtual power plant, the flexible load transfer amount, and the wind and solar power consumption plan are output to form a complete load scheduling strategy that includes scheduling sequence and adjustment actions of each node.

[0170] In one possible implementation, when using the ε-constraint method, load transfer revenue can be set as the primary optimization objective, and upper limits can be set for preset operating costs and carbon emissions. Alternatively, operating costs can be set as the primary optimization objective, and constraint boundaries can be set for carbon emissions and load transfer revenue. By iteratively solving these methods, a load scheduling strategy that satisfies the constraints can be obtained.

[0171] When load transfer revenue is set as the primary optimization objective, the total operating cost cap ε1 and the total carbon emission cap ε2 can be set. By continuously relaxing or tightening the two types of constraint thresholds, the solution set of all non-dominated optimal solutions can be obtained. Then, the optimal dispatch scheme can be selected according to the low-carbon assessment requirements of the power grid, and the optimal dispatch scheme can be used as the load dispatch strategy.

[0172] In this embodiment, the grid operating status, node transferability, and node revenue information are uniformly incorporated into the solution process, so that the scheduling results simultaneously meet the requirements of economy, low carbon emissions, and load transfer revenue, and improve the adaptability of the load scheduling strategy to the fluctuation of new energy output.

[0173] In one possible implementation, the load scheduling optimization method further includes the following steps:

[0174] Step 1: Based on the load scheduling strategy, construct a reward and penalty function that balances scheduling response speed and operational economy.

[0175] For example, based on the load scheduling strategy, a reward and penalty function adapted to dynamic real-time control is built to balance the response time under sudden working conditions with long-term economic and low-carbon comprehensive benefits.

[0176] In one possible implementation, the reward and penalty function consists of two core items: a dispatch speed reward item and a dispatch economic reward item. The dispatch speed reward item incentivizes the algorithm to respond quickly to abnormal conditions such as sudden changes in wind and solar power and sudden load increases, while the dispatch economic reward item measures the cost reduction and carbon emission reduction benefits brought about by load transfer. The priorities of the two items are flexibly adjusted by weight coefficients, with emphasis on response speed during periods of severe grid fluctuations and emphasis on comprehensive benefits during periods of stable operation, providing a unified evaluation standard for reinforcement learning iteration.

[0177] The formula for the reward and punishment function is as follows:

[0178]

[0179] in, This indicates that the scheduling speed reward is negatively correlated with scheduling response time. The economic incentives for dispatching are positively correlated with the reduction in operating costs and carbon emissions. and This represents the weighting coefficient, which is calibrated using historical data to balance the priorities of both.

[0180] Step 2: Use the near-end strategy optimization algorithm to modify and adjust the load scheduling strategy to obtain the adjusted load scheduling strategy.

[0181] For example, the near-end strategy optimization algorithm processes the high-dimensional real-time state of the power grid and continuous power regulation actions. It uses the reward and penalty function as the optimization objective to make real-time dynamic corrections to the load dispatch strategy, thus making up for the shortcomings of mixed integer nonlinear programming, which is only suitable for predicted operating conditions and lagging behind sudden fluctuations.

[0182] In one possible implementation, the near-end policy optimization algorithm limits the single update magnitude of the load dispatching policy through shearing constraints, avoiding drastic fluctuations in regulation actions that could affect grid security; it leverages the policy network to output energy storage and flexible load continuous fine-tuning amounts, and the value network predicts long-term cumulative rewards to continuously iterate and optimize the load dispatching policy.

[0183] For example, the core calculation formula of the near-end policy optimization algorithm is shown below:

[0184]

[0185] in, This indicates the network for which the current policy needs to be updated. This represents the old policy network, with parameters saved before the update, and these parameters remain unchanged during the training process. express Constantly executed control actions express The power grid operating status at all times This represents the ratio of the probabilities of actions taken using the new and old strategies. The advantage function represents the additional reward that the current scheduling action can obtain compared to the average scheduling action. This represents the shearing coefficient, defining the allowable update range for the load dispatching strategy.

[0186] The real-time wind and solar power output deviation, distribution network load, and baseline dispatch instructions of the virtual power plant are concatenated as state inputs. The strategy network outputs load and energy storage continuous adjustment increments. The shear coefficient is used to constrain the differences between the old and new strategies. The network parameters are updated through gradient ascent, and the basic dispatch strategy is corrected in real time.

[0187] In this embodiment, the load dispatching strategy can maintain good dynamic adaptability in scenarios of fluctuating renewable energy output, and make the dispatching results more in line with the comprehensive requirements of distribution network operation constraints and user-side economic constraints.

[0188] In one possible implementation, the load scheduling optimization method further includes: automatically distributing economic compensation to at least one load node to be scheduled using a blockchain smart contract, and using federated learning to protect user behavior data.

[0189] For example, blockchain smart contracts are used to automatically calculate and distribute economic compensation based on the actual amount of load transfer completed, while carbon credits are provided as an environmental incentive. The entire process is conducted without human intervention and the records are tamper-proof. Furthermore, federated learning is used to complete load assessment and scheduling model iteration. The user's original electricity consumption and behavior data are only stored and processed on local edge devices, and only encrypted model parameters are uploaded and aggregated. This avoids the risk of user privacy leakage from the source and forms a complete closed loop of data collection, scheduling execution, incentive settlement, and model iteration.

[0190] In one possible implementation, a unique digital identity is assigned to each load node user on the blockchain. The rules for economic subsidies and carbon credits are fixed within the smart contract. The encrypted summary of the scheduling execution data is uploaded to the blockchain for notarization. A distributed federated learning architecture is built, with the global model distributed to the cloud. Each transformer area trains the model locally using its own user behavior data, and only uploads the model gradient after adding noise and encryption. The cloud aggregates and updates the global model, and the original electricity consumption data does not leave the local terminal.

[0191] Furthermore, a blockchain node can be composed of a dispatch master station, a distribution network operation monitoring unit, a user-side aggregation controller, and verification nodes participating in consensus. The blockchain smart contract pre-writes the dispatch achievement conditions, compensation calculation rules, and settlement confirmation logic. After the load dispatch strategy is generated, data such as node response results and actual power reduction or peak shifting power are written onto the chain. Based on this, the blockchain smart contract automatically calculates the economic compensation to be issued and sends the compensation instruction to the corresponding load node to be dispatched.

[0192] Federated learning allows for the deployment of locally trained models on user-side data collection terminals or edge computing units. These locally trained models can be load prediction sub-models or schedulability identification sub-models built based on user behavior data. After each training round, the parameters are uploaded to the federated aggregation terminal. The federated aggregation terminal weights and aggregates the parameters uploaded from multiple nodes, then distributes the updated global model, which is then used by each node to continue local training. This process iterates the model without sharing the original user behavior data.

[0193] In this embodiment, the combination of automatic settlement of blockchain smart contracts and privacy protection of federated learning enables scheduling incentives to be executed automatically based on on-chain records, and user behavior data to be trained and updated locally. This makes the compensation distribution process traceable, and the data analysis process does not directly expose original sensitive information, thereby improving the data security of the load scheduling strategy.

[0194] The load dispatch optimization method provided in this application first constructs a reinforcement learning reward function based on power grid operating status data and preset dispatch actions. This reward function quantifies the ratio of transfer benefits to transfer costs. Then, based on the load transfer potential assessment value of at least one load node to be dispatched, a Monte Carlo tree search algorithm is used to solve the reinforcement learning reward function to obtain the load transfer value of at least one load node. Finally, based on power grid operating status data, all load transfer potential assessment values, and all load transfer values, a load dispatch strategy is determined. In this embodiment, by jointly modeling power grid operating status data, preset dispatch actions, and load transfer potential assessment values, a reinforcement learning reward function is constructed to reflect the proportional relationship between dispatch benefits and costs. Then, the Monte Carlo tree search algorithm is used to search and evaluate different dispatch paths, thereby outputting load transfer values ​​and forming a load dispatch strategy. This ensures that the load dispatch results are consistent with the current distribution network operating status and forms a more suitable dispatch arrangement for load nodes with higher transfer potential and higher transfer value.

[0195] In one possible implementation, Figure 4 A flowchart illustrating the load scheduling optimization method provided in this application embodiment. Figure 4 .like Figure 4 As shown, the load scheduling optimization method includes the following steps:

[0196] S401. Obtain grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network.

[0197] The power grid operation status data includes node voltage, node current, branch power flow, active power, reactive power, network topology status, distributed generation output, and time-of-use electricity prices. User behavior data includes historical electricity consumption curves, electricity consumption time preferences, interruptible load usage records, and response history.

[0198] S402. Determine the geographical coordinates of each load node to be dispatched and the similarity of electricity consumption change patterns between every two load nodes to be dispatched, and determine the virtual power plant.

[0199] For example, when the distance between the geographical coordinates of two nodes is less than a preset distance threshold and the similarity of their electricity consumption patterns is higher than a set similarity threshold, the load nodes to be scheduled that meet the conditions are clustered into a first virtual power plant. When the load nodes to be scheduled do not meet the clustering conditions, the load nodes to be scheduled are treated as a second virtual power plant. Then, all load nodes to be scheduled are replaced according to the first and second virtual power plants. The subsequent scheduling objects are uniformly represented as the aggregated virtual power plant units or single-node virtual power plant units.

[0200] S403. A pre-set spatiotemporal graph neural network model and a multi-branch long short-term memory autoencoder are used to extract features from power grid operation status data and user behavior data to obtain spatiotemporal feature data.

[0201] For example, a pre-defined spatiotemporal graph neural network model is used to extract spatial features from power grid operation status data and user behavior data to obtain spatially coupled feature data; a multi-branch long short-term memory autoencoder is used to extract temporal features from power grid operation status data and user behavior data to obtain temporal feature data; the spatially coupled feature data and temporal feature data are spliced ​​and fused to obtain spatiotemporal feature data.

[0202] S404. Perform attention-weighted processing on the spatiotemporal characteristic data to determine the load transfer potential assessment value.

[0203] For example, attention-weighted processing is performed on spatiotemporal feature data, and load transfer potential assessment values ​​are calculated based on the weight distribution of each feature dimension, each time slice, and each scheduling object, so as to quantify the transferability of each load node to be scheduled or each virtual power plant unit.

[0204] S405. Based on the power grid operation status data and preset scheduling actions, a reinforcement learning reward function is constructed. Based on the load transfer potential assessment value of at least one load node to be scheduled, the Monte Carlo tree search algorithm is used to solve the reinforcement learning reward function to obtain the load transfer value of at least one load node to be scheduled.

[0205] S406. Based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers, a mixed integer nonlinear programming method is performed to determine the objective optimization function. The ε-constraint method is then used to solve the objective optimization function to obtain the load dispatching strategy.

[0206] For example, the objective optimization function simultaneously takes operating cost, carbon emissions and load transfer benefits as optimization objectives, where integer variables are used to represent scheduling action selection and node start / stop status, continuous variables are used to represent the amount of load transferred, scheduling time and power allocation, and introduces distribution network power flow constraints, node power balance constraints, line capacity constraints, node voltage constraints, user response boundary constraints and load transfer time window constraints.

[0207] Based on the above embodiments, the following are embodiments of the apparatus involved in this application:

[0208] Figure 5 This is a schematic diagram of the load scheduling optimization device provided in an embodiment of this application. Figure 5 As shown, the load scheduling optimization device 500 includes:

[0209] The acquisition module 501 acquires the power grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network;

[0210] The processing module 502 is used to perform feature extraction and attention weighting processing on power grid operation status data and user behavior data to obtain a load transfer potential assessment value of at least one load node to be scheduled, wherein the load transfer potential assessment value is used to quantify the transferability of the load node.

[0211] The determination module 503 is used to determine a load scheduling strategy based on the load transfer potential assessment value of at least one load node to be scheduled, wherein the load scheduling strategy includes: the scheduling order and scheduling action of each load node to be scheduled.

[0212] In one or more embodiments, the processing module 502 is specifically used for:

[0213] A pre-defined spatiotemporal graph neural network model is used to extract spatial features from power grid operation status data and user behavior data to obtain spatially coupled feature data. The pre-defined spatiotemporal graph neural network model includes a feature extraction unit for extracting spatial correlation features and a weight allocation unit for dynamically adjusting the weights of spatial correlation features.

[0214] A multi-branch long short-term memory autoencoder is used to extract time-series features from power grid operation status data and user behavior data to obtain time-series feature data. The multi-branch long short-term memory autoencoder includes: multiple parallel long short-term memory encoders for extracting time-series features of different data respectively and a shared long short-term memory decoder for reconstructing time-series features.

[0215] Spatial coupled feature data and temporal feature data are spliced ​​and fused to obtain spatiotemporal feature data;

[0216] Attention-weighted processing is applied to spatiotemporal characteristic data to determine the load transfer potential assessment value.

[0217] In one or more embodiments, the determining module 503 is specifically used for:

[0218] Based on power grid operation status data and preset scheduling actions, a reinforcement learning reward function is constructed, which is used to quantify the ratio of transfer benefits to transfer costs.

[0219] Based on the load transfer potential assessment value of at least one load node to be scheduled, the Monte Carlo tree search algorithm is used to solve the reinforcement learning reward function to obtain the load transfer value of at least one load node to be scheduled.

[0220] Based on power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers, a load dispatching strategy is determined.

[0221] In one or more embodiments, the determining module 503 determines a load dispatching strategy based on grid operating status data, all load transfer potential assessment values, and all load transfer values, specifically for:

[0222] Based on the power grid operation status data, the assessment values ​​of all load transfer potential, and the values ​​of all load transfers, a mixed integer nonlinear programming is performed to determine the objective optimization function, which takes operating cost, carbon emissions, and load transfer revenue as optimization objectives.

[0223] The objective optimization function is solved using the ε-constraint method to obtain the load scheduling strategy.

[0224] In one or more embodiments, before determining the load scheduling strategy based on the load transfer potential assessment value of at least one load node to be scheduled, the determining module 503 is further configured to:

[0225] Determine the geographical coordinates of each load node to be scheduled and the similarity of electricity consumption patterns between any two load nodes to be scheduled;

[0226] The first load nodes to be dispatched that meet the preset conditions are clustered and aggregated into the first virtual power plant. The preset conditions are: the geographical coordinates are less than the preset distance threshold and the similarity of the electricity consumption change pattern is higher than the set similarity threshold.

[0227] The second load node that does not meet the preset conditions will be treated as a second virtual power plant.

[0228] Replace all load nodes to be dispatched based on the first and second virtual power plants.

[0229] In one or more embodiments, the determining module 503 is further configured to:

[0230] Based on the load scheduling strategy, a reward and penalty function is constructed that takes into account both scheduling response speed and operational economy;

[0231] The load scheduling strategy is modified and adjusted using a near-end strategy optimization algorithm to obtain the adjusted load scheduling strategy.

[0232] In one or more embodiments, the determining module 503 is further configured to:

[0233] The system uses blockchain smart contracts to automatically distribute economic compensation to at least one load node to be scheduled, and employs federated learning to protect user behavior data.

[0234] The load scheduling optimization device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0235] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 600 includes: a processor 601, a memory 602, and a bus 603;

[0236] The memory 602 is used to store the computer-executed instructions of the processor 601;

[0237] The processor 601 is configured to execute the technical solutions of any of the foregoing method embodiments by executing computer execution instructions.

[0238] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0239] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0240] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0241] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0242] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0243] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0244] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0245] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0246] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0247] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0248] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0249] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory (RAM), magnetic disks, or optical disks.

[0250] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0251] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A load scheduling optimization method, characterized in that, include: Acquire grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network; Feature extraction and attention weighting are performed on the power grid operation status data and the user behavior data to obtain the load transfer potential assessment value of the at least one load node to be scheduled. The load transfer potential assessment value is used to quantify the load node's transferability. Based on the load transfer potential assessment value of the at least one load node to be scheduled, a load scheduling strategy is determined, the load scheduling strategy including: the scheduling order and scheduling action of each load node to be scheduled.

2. The method according to claim 1, characterized in that, The step of performing feature extraction and attention-weighted processing on the power grid operation status data and the user behavior data to obtain the load transfer potential assessment value of the at least one load node to be scheduled includes: A preset spatiotemporal graph neural network model is used to extract spatial features from the power grid operation status data and the user behavior data to obtain spatial coupling feature data. The preset spatiotemporal graph neural network model includes: a feature extraction unit for extracting spatial correlation features and a weight allocation unit for dynamically adjusting the weights of spatial correlation features. A multi-branch long short-term memory autoencoder is used to extract time-series features from the power grid operation status data and the user behavior data to obtain time-series feature data. The multi-branch long short-term memory autoencoder includes: multiple parallel long short-term memory encoders for extracting time-series features of different data respectively and a shared long short-term memory decoder for reconstructing time-series features. The spatial coupling feature data and the temporal feature data are spliced ​​and fused to obtain spatiotemporal feature data; Attention-weighted processing is applied to the spatiotemporal feature data to determine the load transfer potential assessment value.

3. The method according to claim 2, characterized in that, The step of determining the load scheduling strategy based on the load transfer potential assessment value of the at least one load node to be scheduled includes: Based on the power grid operation status data and preset scheduling actions, a reinforcement learning reward function is constructed, which is used to quantify the ratio of transfer benefits to transfer costs. Based on the load transfer potential assessment value of the at least one load node to be scheduled, the reinforcement learning reward function is solved using the Monte Carlo tree search algorithm to obtain the load transfer value of the at least one load node to be scheduled. The load dispatching strategy is determined based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers.

4. The method according to claim 3, characterized in that, The step of determining the load dispatch strategy based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers includes: Based on the power grid operation status data, the assessment values ​​of all load transfer potentials, and the values ​​of all load transfers, a mixed integer nonlinear programming method is used to determine the objective optimization function, which takes operating cost, carbon emissions, and load transfer revenue as optimization objectives. The objective optimization function is solved using the ε-constraint method to obtain the load scheduling strategy.

5. The method according to any one of claims 1-4, characterized in that, Before determining the load scheduling strategy based on the load transfer potential assessment value of the at least one load node to be scheduled, the method further includes: Determine the geographical coordinates of each load node to be scheduled and the similarity of electricity consumption patterns between any two load nodes to be scheduled; The first load nodes to be dispatched that meet the preset conditions are clustered and aggregated into the first virtual power plant. The preset conditions are: the geographical coordinates are less than the preset distance threshold and the similarity of the electricity consumption change patterns is higher than the set similarity threshold. The second load node that does not meet the preset conditions will be treated as a second virtual power plant. Replace all load nodes to be scheduled based on the first virtual power plant and the second virtual power plant.

6. The method according to claim 5, characterized in that, The method further includes: Based on the load scheduling strategy, a reward and penalty function that balances scheduling response speed and operational economy is constructed. The load scheduling strategy is modified and adjusted using a near-end strategy optimization algorithm to obtain the adjusted load scheduling strategy.

7. The method according to any one of claims 1-4, characterized in that, The method further includes: The system uses blockchain smart contracts to automatically distribute economic compensation to at least one load node to be scheduled, and employs federated learning to protect the user behavior data.

8. A load dispatching optimization device, characterized in that, include: The acquisition module acquires power grid operation status data and user behavior data of at least one load node to be dispatched in the distribution network; The processing module is used to perform feature extraction and attention weighting processing on the power grid operation status data and the user behavior data to obtain the load transfer potential assessment value of the at least one load node to be scheduled. The load transfer potential assessment value is used to quantify the transferability of the load node. The determination module is used to determine a load scheduling strategy based on the load transfer potential assessment value of the at least one load node to be scheduled. The load scheduling strategy includes the scheduling order and scheduling action of each load node to be scheduled.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images